Towards development of a survival prediction tool for pediatric head injury

The ability to provide an accurate prognosis for children with traumatic brain injury (TBI) would be useful for the children's families and the caregivers. In this study we examined whether an appropriate mathematical model can predict survival in this patient population.

Data from the Children's Hospital of Eastern Ontario (CHEO) TBI registry was analyzed. First, a series of univariate logistic regressions was performed to ascertain the significance of individual predictors, such as age, maximum Glasgow Coma Scale (GCS) score, maximum head injury Abbreviated Injury Scores (AIS) and the Injury Severity Score (ISS). Second, a multinomial logistic regression was fitted using only individually significant predictors and inmodel predictor significance, and interactions were tested. Only two significant predictors were kept in the final model. This final model was subsequently used to predict survival for each individual patient using the n-1 training set (i.e. Lachenbruch's leave-one-out method). The receiver operating characteristics (ROC) method was used to ascertain specificity-sensitivity trade-offs at different probability cut-offs in order to predict survival.

Only the maximum GCS and head injury AIS remained significant, both individually and in the polynomial logistic regression. Empiric ROC curve analyses from leave-one-out survival predictions showed statistical significance (area under the curve = 0.87, Z = 6.8, p < 0.001). Only 12% of cases were misclassified using the 'best' cut-off.

An outcome predictive model for pediatric TBI can be devised using an appropriate mathematical model. It may help to estimate expected outcomes in pediatric TBI more objectively.